dsh-vision-bridge

The official DeepSeek API is text-only — drop a screenshot or a wireframe into DSH's Web UI and the adapter immediately throws UNSUPPORTED_CONTENT, killing the conversation. dsh-vision-bridge closes that gap: it intercepts messages containing image blocks on the request waterfall, hands the ImageBlock to an OpenAI-compatible vision model (DashScope qwen-vl-max by default) to produce a structured Chinese description with verbatim OCR, then re-injects it as a plain text block so DeepSeek receives input it can actually work with.

Beyond passive proxying, the plugin registers two proactive tools — view_image and ocr_image — letting the model inspect local screenshots, diagrams, or design files on demand, or run a dedicated transcription pass. A settings-page card (Vision Model) exposes endpoint, model, and OCR parameters with hot-reload support, so no backend restart is needed. Starting from v0.4.0 the plugin also auto-detects whether the current route natively supports multimodality (e.g. llm-pi-ai); if it does, requests pass straight through with zero overhead — the VL bridge only activates for text-only routes.

It is built for developers who use DeepSeek's text model inside DSH and want to reference images inline during a conversation: debugging screenshots, product mockups, data charts, error screens. Paste them in and the model can describe, analyse, and act on what it sees. The plugin ships with zero hard npm dependencies (core modules are dynamic-imported), does not modify DSH core source code, and runs natively on Windows, Linux, and macOS with no Python, uvx, or WSL requirement.